Point-Cloud Processing with Primitive Shapes
3D acquisition devices usually produce unstructured point-clouds as primary output. A challenge in this context is the decomposition of the point-cloud data into known parts in order to introduce abstractions of the originally unorganized data. This information can be used for compression, recognition and reconstruction. In this project an efficient RANSAC method for automatic detection of primitive shapes (planes, cylinders, cones and tori) in large point-clouds is developed. Building on the decomposed point-cloud several applications are explored. Among these are: (1) The regions of the input point-cloud associated with a primitive shape are resampled as a height field parameterized over the primitive. Image based techniques are used to compress this height information in a way that allows interactive decompression on the GPU during rendering. (2) Recognition of semantic elements, e.g. windows, in the point-cloud data. This is realized with subgraph matching on the primitive's neighborhood graph. (3) Missing parts of the input point-cloud are automatically restored by extending the primitives and resolving the possibly occuring ambiguities in a principled manner. The primtives can also be used to derive an idealized reconstruction.